Chest X-Ray AI Classifier: A Deep

Learning ApproachReporter
 Table of Contents
CONTENTS

1 Introduction to 2 Data Preparation

C h e s t X- Ra y A I and Preprocessing
C l a s s i fi e r

3 M o d e l Tr a i n i n g 4 Flask Application
w i t h Tr a n s f e r Development
Learning

5 Model 6 Challenges,
Performance and Solutions, and
Analysis Lessons Learned
0
1 Introduction to Chest X-Ray AI Classifier
The Objective and Signifi cance
Objective Dataset Origins

To develop and deploy an AI Utilizing a Kaggle-sourced

model capable of dataset consisting of
classifying chest X-ray labeled chest X-ray
images into Normal, images facilitates rigorous
Pneumonia, and training and validation.
Tuberculosis categories.

Early Detection

Early detection of respiratory

diseases can significantly
improve patient outcomes;
reducing healthcare costs and
improving quality of life.
The Context of Respiratory
Diseases

1 2 3

Pneumonia Overview Tuberculosis The Role of AI

Overview

Pneumonia inflames air sacs in one or Tuberculosis, caused by bacteria, AI models offer potential for rapid and
both lungs, caused by infection; can be usually attacks the lungs; can spread if accurate diagnosis, assisting healthcare
milder or life-threatening, especially in untreated, posing significant public professionals in prioritizing and
vulnerable demographics. health challenges globally. managing resources effectively.
0
2 Data Preparation and Preprocessing
 Dataset Organization

Directory Structure Importance of Separation

The dataset is structured into `train`, `val`, and `test` Properly structured directories prevent data
directories for clear segregation of data used for training, leakage, ensuring unbiased model
validation, and testing. performance evaluation and generalization.
 Image Preprocessing Techniques
Grayscale Conversion
Converting images to grayscale reduces
computational load, optimizing training
efficiency without significant loss of crucial
Resizing Images diagnostic information.

Resizing images to 224x224 pixels

standardizes input dimensions, vital
for consistent processing by the
deep learning model.

Normalization
Normalizing pixel values to [0, 1] scales
data, improving model stability and
accelerating convergence during
training.
Addressing Data Challenges
Class Imbalance

Ensuring balanced classes prevents bias towards the majority class, thus creating a
more robust and fair model.

Data Augmentation

Techniques like rotation, flipping, and zooming applied during training augment data,
enhancing the model's generalization capabilities.
0
3 Model Training with Transfer Learning
Model Architecture

1 2

Utilizing MobileNetV2 Transfer Learning

Employing MobileNetV2 as the base model Transfer learning accelerates training, reduces
utilizes transfer learning effectively; data needs, and improves generalization by
leveraging pre-trained knowledge to improve transferring learned features from large
performance. datasets.
 Training Confi guration
Input Shape
Training with (224, 224, 1) for grayscale efficiently matches image
properties to model requirements; optimizing resource usage.

Augmentation Strategy
Data augmentation using rotation, flipping, and zooming enhances
the model's robustness and ability to handle various image
variations.

Optimization Parameters
The Adam optimizer coupled with Categorical Crossentropy loss
facilitates effective model learning and accurate classification.
Training Metrics and Results
Performance Metrics

Achieving 80-90% accuracy on the test set reflects proficient

classification performance validated through rigorous
assessment.

Visual Representation

Graphs showing accuracy and loss over epochs visually

represent training progress, aiding in optimization and
diagnosis of potential issues.
0
4 Flask Application Development
Purpose of the Web Application
User Accessibility

Building a web application allows users to upload

X-ray images, receive predictions, and increases
accessibility to diagnostic aid.

Real-World Applicability

The web application facilitates real-time

predictions; making the AI model practically
useful in various healthcare settings.
Implementation Details
Frontend Design

Using HTML templates with Material-UI enhances user interface design; improving user
experience and interaction with the application.

Backend Logic

The Flask app efficiently handles image uploads and predictions; ensuring smooth
communication between user input and AI classification.
 Addressing
Implementation
Challenges
Shape Mismatch Errors Consistent Preprocessing

Encountering and resolving shape mismatch errors ensures Preprocessing images as grayscale with one
compatibility; aligning model expectations with user- channel ensures consistency between training
provided images. data and real-time inputs; standardizing
processes.
0
5 Model Performance and Analysis
Key Performance Indicators

1 2

Accuracy on Test Set Successful Deployment

Achieving 80-90% accuracy reflects robust Successfully deploying the web app enables
classification capabilities; validated through end-users to test and experience the
rigorous and consistent evaluations. predictive capabilities of the model.
Limitations
Input Channel Limitation

Model trained on 1-channel input constrains the effectiveness; pre-trained weights

designed for RGB images not fully leveraged.

Addressing the Limitation

Retraining with 3-channel RGB inputs may unlock improved performance, leveraging
the full potential of pre-trained weights.
 Sample Predictions

Prediction Display Examples of Output

The Flask App can provide sample predictions to users who Predicted Class: Tuberculosis, Confidence:
upload images, along with a confidence interval. 85.32% might be a typical output to show the
model's potential.
0
6 Challenges, Solutions, and Lessons Learned
Addressing Shape Mismatch

1 2

Problem Identification Solution Strategy

Understanding that shape mismatch inhibits Updating `app.py` to preprocess images

consistent processing is crucial, pinpointing consistently ensures input aligns precisely to
causes in the input pipeline. model expectations; streamlining outputs.
Consistency Is Key
Maintaining Standards

Consistency ensures robust performance

during deployment; vital for producing
accurate and reliable inferences in real
settings.

Pipeline Alignment

Ensuring consistency between training and

inference pipelines eliminates input errors;
promoting reliable classification.
0
7 Next Steps for Model Enhancement
Model Refinement

Retraining with RGB Fine-Tuning

Retraining the model with

01 Fine-tuning the model to
02
3-channel RGB inputs improve tuberculosis
unlocks advantages of pre- detection enhances
trained weights; accuracy; addressing
facilitating performance potential vulnerabilities
improvements and greater and increasing overall
accuracy. reliability.
Deployment Expansion
Cloud Deployment

Deploying the Flask app to a cloud service

broadens accessibility for users; providing
wider utilization and increased impact.

Platform Choice

Heroku and AWS are viable options;

offering scalability to accommodate
increasing use, supporting long-term
viability and wider adoption.
 Evaluation Enhancement
Misclassification Analysis

Analyzing misclassification rates helps identify

patterns; addressing potential gaps while
enhancing model performance.

Targeted Insights

The analysis notebook offers clear insights,

allowing for targeted enhancements and
promoting effective refinements with precision.
0
8 Conclusion and Future Directions
 Project Summary

AI Model Success User-Friendly

Tool

Successfully building Developing a user-

an AI model with 80- friendly Flask app
90% accuracy reflects promotes real-time
meaningful practical predictions; enabling
applicability; faster diagnosis and
confirming viability potentially reducing
for healthcare usage. strain on healthcare
providers.
Future Work

01 02
Enhanced Performance Broad Accessibility

Enhancing model performance through retraining Deploying the app to increase accessibility
increases the model's effectiveness; optimizing makes the tool advantageous for clinical
clinical utility for detecting diseases and purposes; supporting informed decision-making
conditions. and increasing accuracy.
THANK YOU
Reporter